Wind-damaged equipment attached to power lines can cause serious consequences. In extreme cases, that includes igniting wildfires, such as the 2018 Camp Fire. Identifying the structures with hardware most susceptible to wear and tear is the first step in mitigating risks.

PG&E, working with Exponent, developed a machine learning model that could capture at least 80% of worn structures with an 8% precision score. The model combined inspection reports, asset information, site-specific meteorological data, and simulation outputs from a first-principles model of mechanical wear at each of the approximately 148,000 transmission line structures in PG&E’s service area. The model simulated annual wear volume and depth at each structure based on site-specific historical wind data. The dataset included information on structure type, material, and height. Meteorological data provided wind gust, temperature, and humidity.

Key Takeaways:

• Machine learning using first-principles engineering models can help predict mechanical wear on overhead hardware.

• Ensemble tree machine learning models were able to boost prediction accuracy of  mechanical wear across PG&E’s territory and efficiently reduce the manpower needed for lines operation by half.

• Identifying the structures with hardware most at risk of wear and damage may help mitigate the risks of severe consequences such as wildfire ignition.