Predictive analysis of criticality for underground medium voltage lines

In this work, a data-driven method for failure prevention of medium voltage lines is proposed. The main objective is to obtain indications on the grid sections closest to the next fault by analyzing historical data on technical issues and environmental conditions. The proposed approach has been developed by exploiting real data from the distribution network in Umbria (central Italy) provided by e-distribuzione, one of the most important Distribution System Operators in Italy. The proposed analysis focuses on underground electrical infrastructures between primary stations and secondary substations, characterized by a voltage level of 20 kV. This prognostic approach allows the arrangement of preventive maintenance operations by locating the most critical sections of the grid. The proposed method combines statistical information on historical data and machine learning techniques, the former deriving from the know-how of e-distribuzione, and the latter provided by the University of Florence. In particular, classifiers based on Feed-Forward Neural Networks are used to identifywhich underground lines are close to the next fault. The results obtained on the case study considered show the possibility of classifying the time to failure of power lines into short and long periods with an average accuracy level greater than 80 %.