AI and data driven overhead line fault classification for TSO asset management using GIS, weather data, and distance relay measurements
This paper proposes an AI-driven framework that integrates GIS, weather data, and distance relay measurements with modified Long Short-Term Memory (LSTM) neural networks to classify overhead line faults with 83.4% accuracy, thereby enhancing Transmission System Operator asset management and energy system reliability.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Every time a storm sweeps across a landscape, the high-voltage power lines that carry electricity to our homes face a silent, invisible battle. These lines, strung between tall towers across vast distances, are constantly exposed to the elements. When the weather turns violent, the lines can fail, causing blackouts that disrupt daily life and strain the systems that keep the lights on. For the organizations responsible for maintaining these grids, known as transmission system operators, the challenge is not just fixing a broken wire after it snaps, but understanding why it broke in the first place. Was it a sudden gust of wind, a layer of ice, or perhaps a flash of lightning? Knowing the specific cause is crucial for planning repairs and preventing future outages, but pinpointing the exact reason for a failure is often a guessing game, especially when the event happens in a remote area with no one watching.
A team of researchers from the University of Belgrade has developed a new way to solve this puzzle by teaching computers to recognize the fingerprints of different weather-related failures. They combined three distinct types of information: the precise location where a line failed, the weather conditions at that exact spot and time, and the electrical signals recorded by the safety devices that protect the line. By feeding this mixture of data into a sophisticated computer program designed to learn from patterns, they created a system that can look at a record of a power line failure and accurately guess what kind of weather event caused it. This approach moves beyond simple observation, using the history of past failures to build a clearer picture of how the grid reacts to the environment.
The researchers started by gathering a massive amount of real-world data from the Serbian power grid, which spans thousands of kilometers of overhead lines across diverse landscapes. They collected records of over 7,500 specific events that occurred over a ten-year period. For each event, they had to do the hard work of connecting the dots between different databases. First, they took the electrical readings from the safety devices, which told them how far down the line a problem occurred, and translated those numbers into exact geographic coordinates, like latitude and longitude. Once they knew exactly where the trouble happened, they reached back into historical weather archives to pull up the conditions at that specific location during the hours leading up to the failure. They looked at factors like air temperature, wind speed, how hard it was raining, and the amount of moisture in the clouds.
With this rich dataset in hand, the team trained a type of artificial intelligence known as a neural network. You can think of this network as a digital brain that learns by example, much like a student studying flashcards. The computer was shown thousands of past incidents, each labeled with the weather conditions that preceded it and the type of fault that eventually occurred. The goal was for the computer to learn the subtle connections between a specific combination of weather factors and the resulting failure. For instance, the system learned that a certain mix of freezing rain and wind often leads to ice buildup on the wires, while a rapid drop in air pressure combined with high humidity might signal an approaching thunderstorm that could cause a spark. The researchers tested different ways of organizing this information, trying various combinations of weather variables and grouping similar types of failures together to see if the computer could learn faster or more accurately.
The results showed that the computer could successfully identify the cause of a failure in about 83 percent of the cases. This is a significant achievement, as it means the system can correctly diagnose the majority of weather-induced problems without needing a human expert to inspect the site immediately. The study found that the most accurate predictions came when the computer was allowed to look at a window of time leading up to the failure, rather than just a single snapshot of the weather. This allowed the system to see how conditions were changing, such as a sudden spike in wind speed or a steady accumulation of rain, which are often the true triggers for a breakdown. The researchers also discovered that trying to predict every single possible type of failure at once made the task harder, so they found that grouping similar events together helped the computer perform better.
However, the study also revealed the limits of this approach. When the researchers tried to remove the most common type of event from the training data to force the computer to focus only on the rare and difficult cases, the system's performance dropped significantly. This suggests that the computer relies on seeing a wide variety of examples, including the routine ones, to build a solid foundation for understanding the rare ones. Furthermore, the accuracy was not perfect; there were still cases where the computer struggled to distinguish between similar-looking events, such as heavy rain and fog, or where the data itself was unclear. The researchers noted that errors in how the original data was recorded or categorized could confuse the system, leading it to make mistakes.
Despite these challenges, the work offers a powerful new tool for managing the power grid. By automating the process of classifying faults, transmission operators can quickly sort through thousands of past incidents to find patterns that might otherwise go unnoticed. This could help them decide where to reinforce towers, which lines need more frequent inspections, or how to better prepare for specific weather threats. The system is fast enough to run in real-time, meaning it could eventually be used to alert operators the moment a fault occurs, suggesting the likely cause before a repair crew even leaves the station. While the technology is not a magic wand that will eliminate all power outages, it provides a clear, data-driven path toward understanding the complex relationship between the weather and the wires that keep our modern world running. The study confirms that by combining precise location data with detailed weather history, we can teach machines to see the storm behind the spark, turning chaotic events into understandable patterns.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.