A statistically defensible machine learning pipeline for FDIA detection and resilience evaluation in renewable smart grids
False data injection attacks represent a serious cyber-physical security challenge for modern smart grids because falsified measurements can affect state estimation, energy management, and operational decision-making. While various machine learning-based FDIA detection methods have been investigated in literature, many studies report biased results from leakage-prone experiments without rigorous statistical analysis. In this paper, a framework of leakage-safe and physics-informed machine learning for FDIA detection and operational cyber-resilience of renewable smart grids is proposed. The leakage-safe paired normal/attack dataset was constructed using CAISO-derived IEEE 118-bus simulation scenarios integrated with renewable and EMS-related variables. Metadata fields, including scenario ID, sample type, target label, attack severity, and number of attacked buses, were excluded from the model input feature matrix. The framework combines bus-level measurements, grid statistics, physics residual features, and indicators of renewable and EMS and tests Logistic Regression, Random Forest, Extra Trees, HistGradientBoosting, and XGBoost models. The proposed XGBoost model achieved 93.47% accuracy, 93.35% F1-score, 0.9813 ROC-AUC, and 0.9847 PR-AUC on the leakage-safe grouped test set. The repeated grouped split validation indicated good stability with a mean accuracy of 93.81% ± 0.10%. Holm-corrected McNemar tests indicated statistically significant paired-prediction differences between XGBoost and HistGBM, RF, and ET, while the difference between XGBoost and Logistic Regression was not statistically significant. Ablation study, bootstrap confidence intervals, attack severity analysis, and cyber-resilience index provide additional support for evaluation transparency. The results support the feasibility of leakage-safe machine-learning evaluation for FDIA detection under the simulated renewable smart-grid setting.