Prediction of significant wave height using a novel combination of Satin Bowerbird Optimizer (SBO) and light gradient-boosting machine (LightGBM): An explainable paradigm

The precise prediction of the significant wave height (SWH) is the key for marine navigational safety and the maximum possible exploitation of renewable wave energy. Although machine learning (ML) models provides a fast way to make predictions, the current models often have computational inefficiency, hyperparameter optimization, and lack of interpretability. This study fills these gaps by proposing a new, transparent hybrid model that combines the Light Gradient Boosting Machine (LightGBM) and the Satin Bowerbird Optimizer (SBO) to predict SWH at two Canadian coastal stations (South Nomad and Central Dixon Entrance). The choice of LightGBM over alternative models was based on the leaf-wise tree growth scheme of the algorithm, with large datasets of oceanographic measurements of decades handled better by the algorithm in terms of memory consumption than by temporal deep learning, and the SBO algorithm guarantees that global hyperparameters are optimized well. The proposed framework was tested on four meteorological input conditions and strictly compared with classical and state-of-the-art algorithms, including Multi-Layer Perceptron (MLP), Kernel Ridge Regression (KRR), and a SBO-optimized Long Short-Term Memory (LSTM) models. The findings indicated that LightGBM-SBO with the best 7-variable input scenario (Comb4) was highly superior to the benchmark models. It had the best correlation coefficients and minimum errors during the testing period at the South Nomad (R = 0.9538, RMSE = 0.3887 m) and Central Dixon Entrance (R = 0.9168, RMSE = 0.3623 m) stations. In addition, the high reliability of the model was validated by Jackknife min-max uncertainty analysis, which produced narrow intervals (width = 1.982) of prediction with high coverage probabilities (98.6%). To address the black-box constraint, SHapley Additive exPlanations (SHAP) analysis was incorporated, which physically validated the model by defining wind gust speed as the most important meteorological forcing factor. All in all, the proposed framework can be directly mapped into engineering values, ultimately with wave energy flux estimation errors 10–30% lower and a very accurate, interpretable marine decision-making tool.