operational forecasting for an offshore wind turbine: benchmarking data-driven against physics-informed machine learning

Accurate forecasting of operational parameters is essential for predictive maintenance and digital twinning of offshore wind turbines. using a unique dataset from the levenmouth 7mw demonstration turbine, we compare a purely data-driven stacked ensemble model (stackedridge) with a novel physics-informed neural network (get-pinn) that incorporates the energy gradient (k) parameter from energy gradient theory (get). the stackedridge model achieves superior predictive accuracy (rmse = 0.2976, r² = 0.9731) for barometric pressure signals. in contrast, the get-pinn provides valuable physics-aware diagnostics by jointly estimating the flow instability parameter k, supporting the detection of phenomena such as vortex-induced vibrations (viv), albeit with higher forecasting error. these results highlight the complementary strengths of the two approaches: the stacked ensemble for high-fidelity point forecasting and the get-pinn for interpretable, physics-guided maintenance decision support in operational wind farm digital twins.