Confidence-aware GNN for operational reliability assessment under extreme-heat conditions
Year: 2026
Type: article
Abstract: Accurate, real-time assessment of operational reliability under extreme-heat conditions is critical for secure grid operation, yet existing Graph Neural Network (GNN)-based methods often fail to capture scenario shifts and tail dependence between load and wind, resulting in inaccurate estimation of system reliability. This paper proposes a scenario-aware, confidence-calibrated framework that couples a precise load–wind speed modeling block with a probabilistic spatiotemporal GNN. First, days are stratified into normal and extreme-heat regimes. For each regime, per-slot marginal distributions of load and wind speed are selected from a candidate family using K-S/A-D diagnostics, and inter-variable dependence is captured via Student-t, Gaussian, and 270°-rotated Clayton copulas; this design explicitly models high-load/low-wind co-occurrence. Second, a multitask GNN with a shared spatiotemporal backbone and task-specific heads jointly predicts EDNS and LOLF, with MC-Dropout providing calibrated predictive distributions. Evaluations on an RTS-79 test system augmented with wind farms, using a three-year, 15-min multi-city dataset, show that the proposed approach consistently outperforms GCN, GAT, Cheb, and NNConv baselines in mean accuracy (RMSE/MAE) and probabilistic calibration (WIS/CRPS). While richer edge kernels and uncertainty calibration introduce modest inference overhead, the runtime remains practical for day-ahead use. Overall, the framework delivers fast and distribution-robust reliability estimates under heat-induced regime shifts.
Source: Energy Reports
Authors: Bin Wei, Chao Gao, Fan Kuang, Jiajia Huan, Yabin Chen
Language: English
Subfield: Safety, Risk, Reliability and Quality
Field: Engineering
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $3,040
Funder: China Southern Power Grid
Award: G1737139638
OpenAlex ID: W7128318732
