Journal of Energy Research and Reviews

Power systems face climate-driven extremes, compound hazards, rapid electrification and renewable integration, and infrastructure decisions whose consequences persist for decades. These conditions often cannot be represented by historical patterns or well-defined probability distributions. Artificial intelligence (AI) is increasingly applied to forecasting, outage prediction, asset management, optimisation and resilience planning, yet most applications remain prediction-centric, whereas long-horizon grid planning takes place under deep uncertainty. This critical narrative review integrates literature from power-systems engineering, machine learning, climate-risk assessment, infrastructure resilience and decision-making under deep uncertainty to examine how AI can move from forecasting toward robust and adaptive planning. We identify a prediction–planning gap created by two mismatches: a temporal mismatch between the horizons over which AI is predictively competent and those over which infrastructure decisions operate, and an uncertainty mismatch between the stationarity assumptions of learned models and the non-stationary, deeply uncertain conditions of future grids. Distribution shift, rare extremes, model disagreement and optimisation fragility limit prediction-centric approaches. AI offers greater long-term value through scenario exploration, surrogate-accelerated simulation, stress testing, vulnerability discovery, robust multi-objective optimisation and adaptive decision support. We operationalise this shift through an AI-Enabled Adaptive Resilience Planning Framework of six iterative stages, propose a decision-performance evaluation scheme (robustness, regret, adaptability, option value, recovery, equity and tail cost), and set out eight research priorities. Next-generation grid resilience should judge AI not by how accurately it predicts one future, but by how well it supports robust, equitable, explainable and adaptable decisions across many plausible futures.