Explainable machine learning framework for dynamic line rating forecasting with adversarial robustness evaluation

Dynamic Line Rating (DLR) enables more efficient use of existing overhead transmission lines by adapting ampacity to real-time meteorological conditions. While machine learning models have shown promising accuracy for DLR forecasting, their robustness against data perturbations and transparency in decision-making remain critical challenges for practical deployment in power systems. Here we develop an adversarially robust and explainable machine learning framework based on LightGBM for DLR forecasting using real meteorological and transmission line data. Under clean conditions, the model achieves excellent predictive performance with a coefficient of determination of approximately 0.995 and low MAE and RMSE values. Under the simultaneous multi-feature FGSM transfer attack, RMSE increased from 14.86 A to 19.16 A for the conventionally trained model, whereas adversarial training reduced the attacked RMSE to 15.61 A. This result represents empirical robustness under the evaluated transfer-attack protocol rather than certified robustness against adaptive adversaries. Robustness is systematically evaluated under adversarial attacks using the Fast Gradient Sign Method (FGSM) and Basic Iterative Method (BIM). Even under severe perturbations, the model maintains high accuracy with R² values remaining above 0.96. Explainability analysis with LIME and SHAP reveals that wind speed, wind direction, ambient temperature, and solar irradiance are the dominant features, consistent with established thermal principles of overhead lines. The proposed framework uniquely integrates high predictive accuracy, adversarial robustness evaluation, and model interpretability, offering a reliable solution for dynamic line rating in modern power grids facing increasing renewable integration and cyber-physical threats. The proposed framework integrates predictive forecasting, adversarial robustness evaluation, and model interpretability within a unified DLR assessment workflow. Historical lagged meteorological variables together with cyclic temporal encodings enable the model to capture both short-term weather dynamics and seasonal variations affecting conductor ampacity. The resulting predictions are further interpreted using SHAP and LIME to ensure physically consistent decision making.