Abstract
Wave energy offers a scalable, carbon‐free resource but is hindered by costly and logistically challenging in situ measurements. We present a blind-flux, bathymetry-enhanced stacking ensemble to predict WEF along Oman's coast using only remote and reanalysis inputs. Five base learners (RF, XGBoost, CatBoost, MLP, LightGBM) were tuned via randomized search and blended by a Ridge regression meta‐learner. Stacking results were also compared with Support Vector Regression and ElasticNet meta-laerners. On a season‐stratified 80/20 split of hourly records, the ensemble achieves R2 = 0.83, RMSE = 0.151 kW/m, and MAE = 0.112 kW/m, outperforming each individual model by 8–36 %. Feature analysis shows wind drives WEF variability, while bathymetry shapes nearshore patterns. WEF peaks >0.8 kW/m during monsoon, with median ≈0.5 kW/m. Training completes in under 10 min on a standard CPU, far faster than spectral models (SWAN, WW3) and requires no direct energy inputs, making it ideal for data‐poor pre‐feasibility studies. Limitations include extreme‐event underprediction and bathymetry resolution. Future work should integrate directional spectra, physics‐informed models, and high‐resolution depth data. This work demonstrates a pragmatic, accurate, and computationally efficient framework for sustainable wave-energy resource assessment.