Optimization of control-oriented AI models for real-time platform pitch prediction in hybrid wind-wave offshore systems

This study presents a control-oriented deep learning framework for real-time prediction of platform pitch motion in hybrid offshore wind-wave systems. The core contribution lies in a hybrid CNN-LSTM architecture augmented with time–frequency features derived from the fourier synchrosqueezed transform (FSST). This integrated design enables simultaneous spatial pattern recognition, temporal sequence modeling, and spectral enrichment, capturing complex nonlinear dynamics inherent to multivariate time series data. Trained on high-fidelity OpenFAST simulation data, the proposed model achieves a root mean square error (RMSE) of 0.26 degrees with an inference latency of 15 ms, outperforming conventional physics-based and standalone neural models by a factor of 6.4. Robustness is demonstrated across diverse operating conditions, with sensitivity analyses confirming stable predictive performance under varying wave heights, wind speeds, and control inputs. In addition to accuracy, the study addresses deployment feasibility on embedded hardware, model interpretability, and readiness for integration with model predictive control (MPC) architectures. Key limitations, such as generalizability to extreme conditions and explainability, are critically examined, offering a roadmap for future enhancements. By bridging simulation-based learning with real-time control demands, this work advances the applicability of deep learning in offshore renewable energy platforms. It provides a scalable, low-latency, and accurate solution for intelligent vibration mitigation and next-generation control of floating hybrid systems. • OWCs serve as active controllers for hybrid floating platform stability. • Regressive modeling applied to hybrid NREL 5 MW FOWT–OWC system. • Hybrid CNN-LSTM achieves 6.4× accuracy in pitch motion prediction. • FAST validation confirms robustness of control-oriented ML model. • FSST features enhance learning with rich time–frequency dynamics.