Abstract
Reliable medium–long-term wind energy projection is essential in the planning, financing, and operation of large-scale offshore wind development. This study classified projection methods into three categories: statistical/empirical and climate-signal-driven methods, dynamical models with reanalysis and regional downscaling, and machine/deep learning and hybrid methods for bias correction, downscaling, and direct data-driven projection. Then, this study reviewed the technical framework, representative studies, and comparative strengths and limitations. The main finding was that the state of the art increasingly converged on “dynamical simulation plus statistical or machine learning correction”. Next, seven main bottlenecks, along with the countermeasures, were systematically presented: (i) difficult data quality control and insufficient observational representativeness, especially offshore; (ii) divergent, even contradictory, conclusions for the same region across data sources and research groups; (iii) large uncertainty in extrapolating 10 m winds to the continually rising turbine hub height; (iv) difficulty in quantifying and communicating non-stationarity and uncertainty to decision-makers; (v) engineering conversion errors from projected “wind resource” to deliverable “electricity”; (vi) systematic biases in the marine atmospheric boundary layer, strong winds, and extreme conditions; and (vii) unresolved reliability, interpretability, and out-of-distribution generalization of AI models. Correspondingly, three mutually reinforcing strands of countermeasures were proposed: first, strengthening the observational and benchmarking foundation through unified, open, quality-controlled observation networks with data-provenance standards and shared reference datasets and intercomparison protocols; second, advancing physics–data integration and uncertainty quantification through hybrid and physics-informed correction, regime-specific bias correction of boundary-layer and extreme-wind errors, and probabilistic frameworks that delivered and clearly communicated credible intervals; and third, closing the resource-to-electricity gap by embedding power-curve convolution, wake-loss modeling, and availability and technology derating into the projection workflow, with the aim of improving medium–long-term wind energy projection accuracy.