Quantitative AI-Based Load-Flow and Fault-Prediction Modeling for Reliability Enhancement in Electrical Distribution Networks

Abstract

This study examined an integrated AI-based framework for load-flow estimation and fault prediction to enhance reliability in electrical distribution networks. A quantitative design was applied using operational data from 12 medium-voltage feeders over an 18-month period, comprising 104,832 operating snapshots collected at 15-minute resolution and 286 confirmed fault events. Feeder sizes ranged from 48 to 137 buses, with topology depth varying from 6 to 18 levels. Descriptive results showed an overall mean voltage magnitude of 0.984 p.u. (SD = 0.031), with a minimum observed value of 0.901 p.u. and a maximum of 1.067 p.u. Mean feeder loading was 62.8% of rated capacity, with peak loading reaching 118.4%. Outage duration averaged 58.6 minutes (SD = 49.8) and customer interruptions ranged from 14 to 4,320 per event. Load-flow regression models indicated that measurement density significantly reduced voltage prediction error (β = –0.41, p < 0.001), while topology depth increased estimation error (β = 0.28, p = 0.002), producing adjusted R² values of 0.64 for voltage error and 0.58 for branch current error. Fault occurrence modeling demonstrated that environmental disturbance severity (β = 0.44, p < 0.001), electrical stress exposure (β = 0.37, p < 0.001), and asset vulnerability (β = 0.29, p = 0.004) significantly increased fault probability, with Nagelkerke R² = 0.52. Fault location accuracy improved significantly when load-flow-derived stress features were included (β = 0.35, p = 0.001). Reliability-linked regression results showed that improved fault localization reduced restoration time (β = –0.38, p < 0.001), and higher predicted fault probability increased interruption duration (β = 0.26, p = 0.009). Robustness testing under measurement perturbation reduced model explanatory power by less than 6%, confirming stability. Overall, the integrated framework demonstrated statistically significant relationships between state estimation accuracy, fault prediction performance, and reliability outcomes.