Ranking of 6–10 kV Cable Lines to Form a Reconstruction Program Based on Machine Learning Methods

The increase in physical wear of 6–10 kV cable lines makes it relevant to predict failures and plan line reconstruction to ensure the reliability of distribution networks. The purpose of the research is to develop a method for prioritizing cable line reconstruction based on operational data using machine learning algorithms. The study was performed on a dataset of 8198 lines, including failure statistics, general cable characteristics, operating parameters, and test results. The following machine learning algorithms have been tested for forecasting: Logistic Regression, SVM, SVR, RandomForest, GentleBoost, LSBoost, RUSBoost. It is proposed to use two metrics for assessing the quality of the forecast—“Failure capture” and “Selection Precision,” calculated relative to the length of the generated list of cables for re-routing. Due to the fact that these metrics are a compromise (an increase in selection precision may be accompanied by a decrease in failure capture if it is calculated relative to the length), it is proposed to use their harmonic mean when choosing the best model. The Random Forest model showed the best results in terms of forecast quality, making it possible to identify 6.5, 9.7, and 12.6% of future failures over the forecast period of three years when selecting 1, 2, and 3% of the total length of the cable network. The results confirm the effectiveness of machine learning to improve forecasting accuracy and optimize cable line reconstruction programs.