This study applied a Random Forest statistical algorithm to predict failures in underground distribution lines in urban areas, specifically for the cities of Milan and Palermo. In recent years, utilities have observed a significant increase in the frequency of service disruptions on underground electrical lines in urban areas. The availability of failure and electrical load data for Milan, provided by Unareti between 2012 and 2019, enabled the development and optimization of a predictive alert system for these types of disruptions. Crucial to identifying correlations with failure data was also the availability of meteorological data, including atmospheric conditions as well as soil temperature and humidity variables, provided by the meteorological forecasting and reanalysis models used by RSE. The same algorithm was also applied to predict failures in Palermo using historical failure data from e-distribuzione for the period 2010-2019. However, for Palermo, the analysis used Actual Load data for Sicily published by Terna, which can be considered a proxy for load. All variables potentially correlated with failure data were considered, such as the temperature of an ideal underground conductor carrying current, atmospheric variables, and soil temperature and humidity data. For both cities, including events lasting three days provided significant added value in terms of statistical scores; this criterion was also adopted by Unareti for defining a failure event associated with a heatwave. Predictive variable groups and their moving averages from 4 to 20 days were considered to account for the importance of past atmospheric and especially soil conditions in generating failures. For Milan, an optimal system was achieved with groupings of 10 variables and an R² of 0.857, while for Palermo, the R² was 0.798. Generally, for Milan, the algorithm showed greater dependence on variables associated with energy demand, while for Palermo, atmospheric and soil variables had greater statistical significance. The results obtained allow for the configuration of a predictive and alert system for the two cities considered.