Researchers from the Federal Scientific Agroengineering Center VIM of the Russian Academy of Agricultural Sciences together with their independent colleagues from the United States have created a system that uses machine learning methods to predict the likelihood of outages on high-voltage power lines. This tool helps energy companies proactively identify vulnerable grid sections and more effectively plan repairs and upgrades.
Unexpected outages remain a key challenge in the energy sector: according to various estimates, up to half of all grid failures occur on overhead lines. For many years, forecasts had been made on the basis of historical statistics, visual inspections and analysis of weather factors such as thunderstorms, wind, power line icing or fallen trees. For their new study, the scientists took a different approach: they used the internal technical parameters of the lines to identify patterns that are not visible with the traditional approach.
To achieve this, they collected data on 395 power lines 110 kV in the Orel Region and analyzed nine parameters, including line length, proportion of lines passing through forests and populated areas, tower type, service life, technical condition index and wire type. The data were processed via exploratory analysis methods to identify hidden patterns. The processed data were used to train five different machine learning algorithms: support vector machine, logistic regression, random forest and two gradient-boosting frameworks – LightGBM and CatBoost. In order to improve forecast accuracy, the researchers applied special methods to offset sample imbalances, since failures were outnumbered by examples of normal operation.
The best results were demonstrated by logistic regression. This model achieved a prediction quality of 0.78 in training and 0.84 in testing using the ROC AUC metric, which proves that it can confidently distinguish between lines with high and low power outage risks. Moreover, it not only predicted failures but also made it possible to understand the specific parameters creating the risk of an accident.
For instance, it showed that the key risk factor was not a formal assessment of a power line’s condition or its age, but its length: the longer the route, the higher the probability of an accident. Interestingly, power lines passing through populated areas actually reduce the risk, possibly due to more frequent inspections and prompt maintenance. Furthermore, lines with a higher proportion of reinforced concrete towers failed more frequently than those using metal towers. Transit lines also showed a higher accident rate. Meanwhile, parameters such as service life or condition index proved insignificant, likely because critical assets are regularly maintained regardless of age.
Overall, the results of the study clearly demonstrated yet again how machine learning can reveal hidden dependencies and offer a fresh perspective on problems that have been solved with traditional methods for decades. In the future, the researchers plan to expand the database to other regions and time periods so as to create more robust models and move closer to the practical implementation of similar forecasts at energy companies.
