Data-Driven Prediction of Wave Energy Potential in the Black Sea Using Ensemble Learning and Real-Time Buoy Observations

Abstract This study proposes a machine learning-based framework for short- to medium-term (1–72 h) wave energy forecasting using buoy-measured wave parameters collected from Samsun and Ordu buoys in the Black Sea. Multiple machine learning (ML) techniques, including Long Short-Term Memory (LSTM), Extreme Learning Machine (ELM), Support Vector Machine (SVM), Bayesian Regression, Residual Neural Networks (ResNet), and Recurrent Neural Networks (RNN), were implemented to estimate wave power using forecasted wave parameters. Each model produced independent predictions, which were then integrated through a stacking ensemble approach based on Extreme Gradient Boosting (XGB) to obtain the final wave energy estimates. When compared individually, the ensemble framework provided noticeably lower error levels across all performance indicators, including RMSE, MAE, and MSE, and achieved higher coefficients of determination. Among the tested approaches, the XGB-based ensemble showed mostly reliable performance, particularly for short-term forecasts. For one-hour-ahead predictions, RMSE values were generally between 0.01 and 0.02 m, while R 2 values were consistently above 0.99. Although prediction accuracy gradually decreased for longer horizons, the model remained stable for lead times of up to 72 h. Differences were also observed between buoy locations. Predictions derived from the Samsun buoy were more stable than those from Ordu, which is likely related to local variability in wave conditions. Our findings suggested that site-specific dynamics play an important role in wave energy forecasting. Overall, this work demonstrates the potential of combining real-time buoy measurements with ensemble machine learning to estimate wave energy in the Black Sea, providing a practical alternative to computationally intensive numerical wave models and supporting future offshore energy planning.