A causal-inference-fused model for automatic identification of outage causes in distribution networks

Abstract

Automatic outage-cause identification in distribution networks is often learned from historical records in which feeder zone, season, and sensor quality are spuriously correlated with the recorded cause, so performance can deteriorate when operating conditions change. This paper develops a causal-inference-fused (CI-Fusion) model that combines a structural causal graph, graph-path evidence construction, normalized inverse-propensity weighting, counterfactual measurement augmentation, and calibrated fusion of linear and boosted-tree probability heads. The model uses electrical responses, weather triggers, asset condition, relay logs, infrared/visible inspection evidence, and fiber-temperature gradients, while coarse historical priors are used only for training balance and are excluded from inference. Evaluation under context shift, label uncertainty, sensor noise, and missing channels showed an out-of-distribution (OOD) accuracy of 90.07%, a macro-averaged F1 score (macro-F1) of 89.86%, and a macro-averaged area under the receiver operating characteristic curve (macro-AUROC) of 98.12%. The macro-F1 exceeded that of Extreme Gradient Boosting (XGBoost) by 2.23 percentage points, and remained 82.09% under Gaussian noise combined with 20% missing evidence, 4.42 points above the strongest baseline. Per-record batched central processing unit (CPU) inference required 0.0029 ms, excluding data acquisition. By separating physical evidence from unstable historical context, this study addresses a central barrier to dependable automated diagnosis. Its potential impact is safer, more transparent, and more transferable fault isolation, inspection prioritization, and restoration decision support across changing distribution-grid regimes.