Physics-informed neural network-based dynamic line rating forecasting under cyberattacks

This paper presents a physics-informed neural network (PINN) framework for forecasting dynamic line rating (DLR), enhancing prediction accuracy, physical interpretability, and resilience to cyber-induced input manipulation. Unlike traditional data-driven models, the proposed PINN integrates thermal balance equations of transmission lines directly into the learning architecture, ensuring physically consistent outputs. The model is trained using meteorological data from a 400 kV transmission corridor connecting two wind farms. Experimental results demonstrate that the PINN model outperforms other models, achieving RMSE values of 2.62, 5.66, and 6.31 for 6-hour, 3-month, and 6-month forecasting horizons, respectively. Compared with models such as Bi-LSTM and SVR, the PINN model shows significant improvements in accuracy. Moreover, under cyber-scaling attacks that manipulate input weather data, the PINN retains stable performance, with minimal degradation in forecasting accuracy. These results underscore the critical advantages of PINNs for DLR forecasting, offering robust, interpretable solutions for grid operations under cyber-physical threats.