Researchers in Turkey and Ukraine have trained a machine-learning system to forecast how much current an overhead power line can safely carry when its weather data have been deliberately altered. In tests reported by Scientific Reports on September 11, the protected model held its forecast error to 15.61 amperes, compared with 19.16 amperes for a conventionally trained model under the same simulated data attack.
Power lines heat up as current passes through them. Excess heat can stretch a conductor and make it sag, so grid operators assign each line a safe current limit, known as its ampacity. A static rating fixes that limit using conservative weather assumptions. Dynamic line rating updates it from conditions such as air temperature, wind and solar heating, allowing operators to use available capacity without exceeding the line's thermal limit.
The new system uses LightGBM, a machine-learning method built from decision trees. It combines current and earlier weather readings with time-of-day and seasonal information, then predicts ampacity in amperes. With unaltered inputs, the model recorded a coefficient of determination of about 0.995 and a root mean square error of 14.86 amperes.
The researchers then introduced small, coordinated changes to several input values using two established attack methods. These tests mimic a compromised or corrupted data feed rather than an attack on a live grid. The conventional model's error rose from 14.86 to 19.16 amperes in one multi-input test. A model trained on perturbed examples recorded 15.61 amperes under that test, and the reported coefficient of determination remained above 0.96 under the more severe perturbations evaluated.
The team also traced which inputs drove individual forecasts. Wind speed, wind direction, ambient temperature and solar irradiance carried the most influence, matching the physical processes that cool or heat a conductor. That agreement gives an operator a way to question a forecast whose explanation conflicts with the weather around the line.
The study was an offline evaluation of one forecasting workflow, not a field trial or a security guarantee. Its authors state that the result demonstrates resistance to the tested transfer attacks, not protection against an attacker who adapts to the deployed model. The next test is a live operating environment with sensor faults, communication outages and adaptive attacks, together with a verified fallback to a conservative line rating when the forecast cannot be trusted.
Sources
- Feyyaz Alpsalaz and colleagues, Scientific Reports, September 11, 2026. DOI: 10.1038/s41598-026-69917-1. Forecasting method, attack tests, results and limitations.
- US Department of Energy, Advanced Transmission Technologies report. Dynamic line-rating operation, weather inputs and grid-control safeguards.
- Lead image: “Electricity Pylons” by Geographer, used under the CC BY-SA 2.0 licence.
