| Accurately predicting the duration of weather-induced power outages is essential for improving grid resilience and emergency response. However, conventional evaluation metrics such as MAE, MSE, and macro-F1 fail to account for the varying costs of misclassifications. This paper introduces a cost-aware macro-F1 metric, a flexible extension of the standard F1 that incorporates misclassification severity through a customized cost matrix. The proposed metric operates directly on the confusion matrix, allowing practitioners to align evaluation with real-world costs without modifying the training procedure. Using a large-scale outage dataset and synthetic confusion matrices, we demonstrate that models with identical macro-F1 can exhibit substantially different cost-aware performances. The metric consistently reflects the severity and direction of ordinal errors, providing deeper insight into model behavior. Results across multiple temporal forecasting settings confirm its robustness and adaptability, paving the way toward cost-sensitive evaluation in resilience analytics. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.