Robustness in AI-Based Power System Forecasting: A Structured Review and Unified Evaluation Framework

Artificial intelligence (AI) has advanced load, solar PV, wind, and net-load forecasting, yet evaluation remains largely accuracy-centric, with limited assessment of reliability under missing data, measurement perturbations, distribution shifts, extreme events, and training variability. This paper presents a structured review and a proposed standardized robustness evaluation framework for AI-based power-system forecasting. Robustness failures are organized into four dimensions: data imperfections, distributional non-stationarity, extreme/out-of-distribution (OOD) conditions, and intrinsic model instability. The resulting failure-mode-driven framework integrates tail-error and regime-aware analysis, probabilistic calibration, controlled stress testing, cross-domain validation, and retraining-stability assessment. A preliminary benchmark across three datasets, two forecasting tasks, three model families, two forecast horizons, and five independent training runs per configuration demonstrates that nominal accuracy and robustness criteria can favor different models. Robustness-enhancing methods are further organized at the data, model, and training levels, and their implications for reserve scheduling, risk-aware operation, and economic performance are examined. The review highlights persistent gaps in calibration assessment, systematic stress testing, cross-domain evaluation, and reproducible benchmarking. These findings motivate a shift from accuracy-centric evaluation toward multidimensional, deployment-oriented robustness assessment for trustworthy AI-based power-system forecasting.