Machine-Learning-Based Intelligent Safety Risk Early Warning for Live-Line Work in Distribution Networks

Abstract

Live-line work reduces scheduled outages and improves distribution reliability, but it also places workers in a changing environment where energized parts, tool condition, human behavior, weather, and work-space constraints interact. Conventional controls based on work-ticket review, manual supervision, and fixed thresholds are effective for explicit prohibitions but are less capable of recognizing a gradual rise in risk. This study develops a multisource data system covering work plans, personnel qualifications, tool tests, grid status, environmental sensing, video events, and historical incidents. It proposes an early-warning framework that combines rule constraints, machine-learning estimation, probability calibration, graded response, and feedback-based updating. XGBoost is used for structured records, while temporal and visual modules capture state trends and unsafe behavior. Their calibrated outputs are fused with mandatory rules rather than allowed to override them. Cost-sensitive learning, task-and-time-grouped validation, calibration analysis, and interpretable evidence are included to address rare hazardous events, leakage, false alarms, and opaque decisions. A four-level response process, human confirmation, drift monitoring, and model audit are then specified. The framework is intended to support earlier intervention while retaining the authority of safety codes and the decision responsibility of on-site personnel.